Frontier Log

Benchmark · Active

ARC-AGI-3

Created by ARC Prize Foundation, 25 Mar 2026 · Higher is better · Website

0102030Jul 202613.3%GPT-5.6 Sol0102030Jul 202613.3%GPT-5.6 Sol

Settings and run setups

Task setProviders
Semi-Private set (55 environments)OpenAIAnthropic
Public set (25 environments)OpenAI

Download image

Results

ModelScoreSource
GPT-5.6 Sol9 Jul 202613.3%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkCorroborated The same number is printed in 1 other lab document: 1official harness · max reasoning effort · reported laterOur capture
GPT-5.6 Sol9 Jul 20260.9%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.official harness · low reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 202638.3%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkCorroborated The same number is printed in 1 other lab document: 1harness with retained reasoning and compaction · max reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 20261.5%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.official harness · medium reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 20265.2%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.official harness · high reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 20267.2%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.official harness · xhigh reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 20263.7%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.harness with retained reasoning and compaction · low reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 20267.3%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.harness with retained reasoning and compaction · medium reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 202613.4%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.harness with retained reasoning and compaction · high reasoning effort · reported later · another report, not chartedOur capture
GPT-5.6 Sol9 Jul 202625.7%How enabling two settings tripled our scores on the ARC-AGI-3 benchmarkChecked A person checked this number against its source on 11 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.harness with retained reasoning and compaction · xhigh reasoning effort · reported later · another report, not chartedOur capture

All benchmarks